⚡ The Prompt Engine — Uncensored (OpenRouter & Ollama Edition)

Deutsche Version

Functional rebuild of thepromptengine.midnightlabai.com — without its fatal flaw: the original hardcoded gemini-2.5-flash in its frontend bundle; after Google retired that model for new API accounts, every new key ran into a 404.

This clone uses freely selectable vision models instead — no model is hardcoded, and the model list is loaded at runtime. Switchable between two providers:

• OpenRouter (cloud) — huge model catalog, API key required, costs per request.
• Ollama (local) — runs on your own machine: no API key, no cost, and no image ever leaves your computer.

The switch sits in the top-right header (☁️ OpenRouter / 🖥️ Ollama) and in ⚙️ Settings. Model, API key, and Ollama address are stored per provider, so switching back and forth never loses a setting.

Features

• Image → Prompt: Upload an image (drag & drop, click, or Ctrl+V); a vision LLM analyzes it and produces a detailed generation prompt.
• Idea → Prompt: Enter a short idea and the engine expands it into a professional prompt.
• Image → Image: Upload a source image and describe in words what should change — out comes a ready-to-use edit prompt for image-editing models. The change scope (minimal / moderate / strong) controls how much may change alongside the requested edit; on "minimal" the engine spells out what has to be preserved (identity, pose, background, lighting, framing) in the prompt itself.
• Image → Video: Upload a start image, optionally an end frame as the target frame, and describe what should happen in the clip — out comes an image-to-video prompt. Also selectable: camera movement (dolly, pan, orbit, crane, handheld …), motion intensity (subtle / moderate / dynamic) and clip length.
• Target platforms, per mode:
  • Image/Idea → Prompt: SDXL (tags + negative prompt), Pony/Illustrious (booru tags incl. score/rating tags), Flux (prose), Midjourney (incl. --ar/--stylize), DALL-E, Ideogram, Nano Banana/Gemini, Qwen-Image (with text rendering), Krea 2 (photorealistic), Z-Image Turbo (no negative prompt), universal.
  • Image → Image: Nano Banana/Gemini, Flux.1 Kontext, Qwen-Image-Edit, Seedream/SeedEdit, GPT-Image/DALL-E edit, SD/SDXL img2img + inpainting (with a denoise recommendation and the region to mask), universal.
  • Image → Video: Kling, Runway Gen-4, Google Veo 3 (with audio description), Hailuo/MiniMax, Luma Dream Machine/Ray, Wan 2.2, Sora 2, Seedance, universal.
• Image formats: Aspect-ratio selection as in the original (Auto, 1:1, 16:9, 9:16, 4:3, 3:2) — emitted as --ar for Midjourney, with a matching resolution recommendation for SDXL/Pony. Image → Image drops the choice: there the source image dictates the format.
• Technical parameters (as in the original, collapsible): Logic Mode, Camera Angle, Shot Type, Perspective, Composition, Lighting, Atmosphere, Mood, Emotion — each with the same value lists as Midnight LAB v2.0.
• Bilingual: Language switch (DE/EN) in the top right; the choice is remembered.
• Options: Style preset, detail level, 3-variation mode, streaming output, copy button, local history.
• Send to ComfyUI (LoRA Manager): The paper-plane button next to “Copy” puts the prompt and negative prompt straight into text nodes of the workflow open in ComfyUI — through the interface of ComfyUI-Lora-Manager. See Send to ComfyUI.
• Uncensored: NSFW toggle — explicit images are analyzed directly and without euphemisms. The recommended default models (Qwen3-VL family) run on unmoderated OpenRouter endpoints; running locally via Ollama, only the model itself decides. Hard limits always stay active: no depictions of minors, no sexual content involving real, identifiable people.
• No backend: Pure static frontend. The API key stays in the browser's localStorage and is sent exclusively and directly to openrouter.ai — or, with Ollama, to your own machine.

Screenshots

The four modes in the interface. Source images, inputs, and results are illustrative examples — in actual use the prompts come from the selected model.

🖼️ Image → Prompt

Upload an image, optionally add an extra instruction — out comes a generation prompt in the format of the chosen target platform (SDXL with a negative prompt here).

💡 Idea → Prompt

No image needed: a short idea is enough, and the engine expands it into the full prompt. The image area disappears; every other option stays.

🎨 Image → Image

A source image plus a change description yields an edit prompt. This mode adds the change scope; the aspect-ratio selection is gone because the source image dictates the format.

🎬 Image → Video

Start image, optional end frame as the target frame, and the video description — plus camera movement, motion intensity, and clip length.

Usage

1. Start the page:

   • Windows: double-click start.bat. It checks whether Ollama is running and starts it if needed, picks a free port, starts the server in the right folder, and opens the browser.
   • macOS / Linux: python3 serve.py in the project folder — same thing.
   • To skip the Ollama check: start.bat --no-ollama / python3 serve.py --no-ollama.
   • By hand: python3 -m http.server 8080 in the folder holding index.html, then open http://localhost:8080.

   For Ollama the page has to come from localhost. Hosting it on GitHub Pages / Netlify works with OpenRouter only — see Option B.

2. Pick a provider in the top-right header and set it up under ⚙️ Settings (see below).

3. Pick a mode at the top left, upload an image or enter an idea → ⚡ Generate prompt.

Option A: OpenRouter (cloud)

1. Create an API key at openrouter.ai/keys.
2. Enter the key under ⚙️ Settings and pick a vision model.

Option B: Ollama (local)

1. Install Ollama and pull a vision model:

   ollama pull qwen2.5vl:7b

2. Start Ollama (ollama serve, or just leave the app/service running).
3. Open this page over http://localhost — easiest via start.bat (Windows) or python3 serve.py (macOS/Linux).
4. Switch to 🖥️ Ollama in the header. ⚙️ Settings holds the address (default http://localhost:11434); 🔄 Load model list shows every installed model, split into "with image support" and "text only".

That's all: Ollama already allows requests from localhost, 127.0.0.1 and 0.0.0.0 on any port out of the box. OLLAMA_ORIGINS is only needed when the page runs at some other address — see below.

When the page isn't served from localhost

• Served over https (e.g. GitHub Pages): the browser blocks the call to http://localhost:11434 as mixed content before Ollama is ever asked. No Ollama setting fixes this — the page has to run locally over http://localhost.
• Opened by double-click (file://): the browser sends origin null, which Ollama rejects. Again: serve it from a local server.
• Some other address, e.g. a LAN IP: only then does the origin need to be allowed explicitly:

  OLLAMA_ORIGINS='http://192.168.1.50:8080' ollama serve

Setting environment variables persistently

They must reach the server process, not the shell you run ollama run in — restart Ollama afterwards:

# Windows: quit Ollama from the tray icon, then
setx OLLAMA_CONTEXT_LENGTH "8192"

# macOS (app): does not survive a reboot
launchctl setenv OLLAMA_CONTEXT_LENGTH "8192"

# Linux (systemd): sudo systemctl edit ollama.service
[Service]
Environment="OLLAMA_CONTEXT_LENGTH=8192"

Further notes

• Image → Prompt, Image → Image and Image → Video need a multimodal model (qwen2.5vl, llava, minicpm-v, llama3.2-vision, gemma3). Text-only models work in Idea → Prompt mode only.
• The optional end frame in Image → Video sends two images in one request. Not every model handles multiple images equally well; if the model ignores the second one, drop the end frame and state the target state in the video description instead.
• Keep OLLAMA_CONTEXT_LENGTH at 8192 or above: image tokens plus the long system prompt otherwise overflow the default context window and the start of the prompt gets truncated.
• If access still fails, the error message in ⚙️ Settings names the actual cause (mixed content, file://, foreign origin, or service unreachable). Note that opening http://localhost:11434 in the address bar proves nothing — that request carries no origin for the browser to check.

Send to ComfyUI (LoRA Manager)

This needs ComfyUI-Lora-Manager installed as a custom node in ComfyUI. Its browser extension reports the nodes of the open workflow to its backend, and that is exactly how The Prompt Engine puts the text in — the same route the “Send to workflow” button in LoRA Manager takes.

1. Start ComfyUI and keep the workflow open in the browser.
2. Generate a prompt, then click the paper-plane button LoRA Manager next to “Copy”.
3. In the dialog, check the ComfyUI address (default http://127.0.0.1:8188) and pick one target node each for prompt and negative prompt. With three variations, pick the variant as well. Send replaces the node's text — or appends to it when the box is ticked.

Offered as targets are active nodes of type Prompt (LoraManager), Text (LoraManager) and CLIP Text Encode whose text field isn't fed by a link from another node. Any other text node can be enabled in ComfyUI via right-click → Mark as → Send Prompt Target. The page remembers the nodes you picked last; without history, a node with “neg” in its title gets the negative prompt.

Connection: ComfyUI rejects requests from pages on another port (HTTP 403) unless it runs with --enable-cors-header. That is why start.bat and serve.py include a small proxy that forwards only these two calls to ComfyUI and only accepts requests from this page itself — so ComfyUI's default configuration is enough. If the page is served some other way (e.g. with python3 -m http.server), the request goes to ComfyUI directly, and ComfyUI has to be started with --enable-cors-header. From an https page (GitHub Pages), an http ComfyUI is unreachable because of mixed content.

Model notes

• OpenRouter, for uncensored analysis: Qwen3 VL 235B / 30B / 32B or Qwen2.5 VL 72B (unmoderated endpoints, cheap). Grok is also fairly permissive.
• Ollama, local: qwen2.5vl:7b (~6 GB, fast, barely moderated), qwen2.5vl:32b (best quality, ~21 GB VRAM), llava:13b, minicpm-v:8b (lightweight). llama3.2-vision and gemma3 work well but are noticeably more moderated.
• GPT, Claude, and Gemini models work for SFW content but usually refuse NSFW images.
• If a model ever becomes unavailable (the original's bug), just pick another one in the settings — or enter any model ID by hand.

Legal

For adult users only. When using the cloud, the terms of use of OpenRouter and of the respective model provider continue to apply and are the user's responsibility. When running locally via Ollama, the licenses of the pulled models apply.